4 A 21.5M-query-vectors/s 3.37nJ/vector reconfigurable k-nearest-neighbor accelerator with adaptive precision in 14nm tri-gate CMOS

Energy-efficient k-nearest-neighbor (kNN) computations are key building blocks for computer vision, classification, and machine-learning workloads 1-3. Determining distances to high-dimensional vectors within a large vector database results in high compute cost. Adaptive precision improves energy efficiency by eliminating a majority of vectors without costly full-precision computation, with as-needed precision refinement to guarantee kNN accuracy of closely matched vectors. A special-purpose on-die kNN accelerator with 128-dimensions by 128 parallel reference vectors, targeted across mobile SoCs to multi-core microprocessors, and reconfigurable for either Manhattan or Euclidean distance, is fabricated in 14nm tri-gate CMOS 6. Partial distance compute circuits, 2b window-based sort, MSB-to-LSB-based selective distance refinement, robust ultra-low voltage circuits, and state tracking control to selectively resume next-nearest candidates enable nominal energy efficiency of 3.37nJ/query vector or 9.7TOPS/W (measured for 21.5M vectors/s, 16 cycles/vector at 750mV, 25°C) with a dense layout occupying 0.333mm2 (Fig. 14.4.7) while achieving: i) scalable performance up to 26.4M vectors/s, 114mW measured at 850mV, ii) 2-cycle latency and 43pJ energy to find each subsequent nearest neighbor, iii) up to 5.2× higher throughput while maintaining full-precision kNN accuracy, iv) 16× search-space reduction for next-nearest neighbor, v) ultra-low voltage operation measured at 360mV, 1.1M vectors/s, 1.44mW, and vi) peak energy efficiency of 1.23nJ/vector at 390mV (near-threshold), 25°C.